The undefeatable bot, and a problem

#1 Yes...

#2 An unbeatable bot is not too hard to write.

A bot that's fun to play against is much harder. :)

#3 Ok then if a undefeatable bot is easy to make, can you make a bot that uses the term 'monkey see monkey do' literally (oh and make it remember what it sees, so it will reuse tactics that work)

#4 Making an undefeatable bot is simply a matter of making it never miss and always dodge. Making a bot that watches and learns is something the entire scientific community hasn't been able to accomplish yet. And it would most likely require more processing power and available memory than any household computer.

#5 That's not true at all, Rockatansky. Ever played Black & White?

AI has got a whole sub-field called 'Machine Learning' that's dedicated to this type of question, and the researchers there _do_ make progress. It's not like these researchers have been butting their heads into the wall all these years.

#6 I'll give you an example... there is a program recently introduced to the cinematic special effects market called Natural Motion that uses a newly developed technology called artificial evolution to do something similar to what you want.

http://www.naturalmotion.com/pages/technology_hiw.htm

The artificial evolution logic can teach a bipedal 3d model how to walk and run without programming a single frame of animation. You just tell it where to go and how fast to get there and it learns to walk!

Natural Motion is merely a tool that takes advantage of a portion of this technology to produce realistic animations for cg actors.

It retails for $16,000...

#7 Black & White's "intelligent" monsters are nothing more than a series of variables (chance to kill, chance to eat, etc) that are altered every time you punish or reward him based on the last action he performed.

#8 To utilize that in the game you would need to program a series of tactics the bot would perform, and alter the chance of him using those tactics based on the success of the last tactic used. This would make him appear to learn, but he would never be able to perform anything new, you would have to reprogram him for every new tactic you wanted him to employ.

#9 Yes. A neural network can be described as a 'series of variables' but that's grossly oversimplifying the approach.

You make it sound like they used some sort of switch statement - they didn't.

#10 Of course is simplifying it, it's probably a billion lines of code, but the overall structure is just that!

If you have a bot in a torque game, and you give it a few unique variables (chance to jump, chance to run, chance to shoot) and every time it does one of those things and dies, the chance variable is lowered, then it will be less likely to do that in the future. That's incredibly simplistic, but I believe that's how the overall structure works.

#11 That's like saying the overall structure of a computer game is that it moves pixels around based on user input. ;)

Here's a nice introductory article on neural nets: www.ai-junkie.com/nnt1.html

#12 Who cares what learning technology it used? Black and White is abominably awful. And they're making another one! For God's sake!

#13 Yes, this issue is stupid, overall, to do something like you said depends on the way you do it.

#14 Well cant you have a bot that if killed by a guy with a machine gun, it would be more likely to use a machine gun. and why cant it be programed to remember the last couple of moves the player made befor it killed the bot? then repeat the same manuvers?

#15 Of course you can do that.

#16 Look this is a really stupid issue.

If you have a game where winning depends solely on finding the square root of a very large number in less than a second, you can write a "bot" to beat a human at it every time.

If you have a game where winning depends on identifying a letter in a very garbled picture in less than a second, you probably can't write a bot to beat a human every time.

In an fps, how the hell do you define "cheating" anyway? I mean, sure, doing double the damage of a human is cheating, but a bot can probably aim better just by having a different set of information, which is cheating in itself.

And what the hell would the use of remembering the last couple of "moves" a human made before it was killed be? Those moves would probably be completely different if both players were only like 10 cm away from the original positioning.

#17 One thing one might consider is altering your graph connectivity info based on "observed" player movements. That way the AI could learn common player used routes (either to use them itself or avoid them). Or if it sees someone getting fragged, backing off a bit. That sort of thing.

#18 Tommy Hanusa wrote:
"Could you make an undefeatable bot? im talking about a bot that has the same advantages as a player, the move as fast, shoot as fast, have the same health, weapons,and weaknesses, but is undefeatable, in every situation. the undefeatable bot could shoot,snipe,dodge,and pilot better than any other human or bot."

Your description asks for a BOT that is such a good player that he is unbeatable ( Does not use the fact that it has access to the game engine and could win by "cheating" ).

If so: the answer is probably no ( for at least 50 more years that is ). You could use advanced AI techniques to make a BOT that is very hard to defeat, but not "undefeatable".

#19 Um just give the bot infinite health!

#20 If you made the bot get a head shot as soon as it layed eyes on you and dodge grenades it would sure be invincible. Like the agents in the matrix they could do things better than any human and had never been defeated (neo being an exception) you would make a bot like that and it would most likely be unbeatable. I was actually thinking of making an invincible ally, that would save a lot of screaming at the idiot computer that grenades itself and has trouble hitting with a heatseeking rocket. God I hate those things.

P.S.
he said it would have to have the same phsical attributes as the human so it couldn't have infinite health.